llava/mtmd : fixes to fully support dl backends (#13303)

tools/CMakeLists.txt

@@ -27,11 +27,11 @@ else()
     add_subdirectory(run)
     add_subdirectory(tokenize)
     add_subdirectory(tts)
+    add_subdirectory(llava)
     if (NOT GGML_BACKEND_DL)
         # these examples use the backends directly and cannot be built with dynamic loading
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
-        add_subdirectory(llava)
         if (GGML_RPC)
             add_subdirectory(rpc)
         endif()

tools/llava/clip.cpp

@@ -3382,7 +3382,15 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             GGML_ABORT("Unknown projector type");
     }
 
-    ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
+    // ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
+    ggml_backend_dev_t dev = ggml_backend_get_device(ctx->backend_cpu);
+    ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
+    if (reg) {
+        auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
+        if (ggml_backend_set_n_threads_fn) {
+            ggml_backend_set_n_threads_fn(ctx->backend_cpu, n_threads);
+        }
+    }
 
     auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf);
     if (status != GGML_STATUS_SUCCESS) {

tools/llava/llava.cpp

@@ -2,6 +2,7 @@
 #include "llava.h"
 
 #include "llama.h"
+#include "ggml-cpp.h"
 
 #include <algorithm>
 #include <cerrno>
@@ -209,7 +210,10 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
     // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
     ggml_build_forward_expand(gf, flatten);
-    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+
+    ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) };
+    ggml_backend_graph_compute(backend.get(), gf);
+
     struct ggml_tensor* result = ggml_graph_node(gf, -1);
 
     memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context